Initial release: code for WCL2026-1544 (context-aware embedding masking via DRL)

run_sweeps.sh

@@ -0,0 +1,65 @@
+#!/bin/bash
+# Batch runner for U-sweep and ablations on real BERT embeddings.
+# Each run uses 60 epochs x 150 steps (~5 min on MPS).
+# Outputs saved under results_drl/ with per-run subdirs.
+
+set -e
+
+BASE="/Users/kyo/Documents/AY/논문/5. WCL-DRL"
+cd "$BASE"
+
+EMB="bert_agnews_8000.pt"
+EPOCHS=60
+STEPS=150
+PY="/opt/anaconda3/bin/python"
+
+mkdir -p results_sweeps
+
+echo "=== U-sweep: DRL ==="
+for U in 1 2 3 5 6; do
+    echo ">>> DRL U=$U"
+    $PY Code/drl_mask_policy.py \
+        --mode drl --embed-file $EMB \
+        --users $U --mux-factor 4 \
+        --epochs $EPOCHS --steps-per-epoch $STEPS \
+        --save-dir results_sweeps/drl_U${U} \
+        2>&1 | tail -3
+done
+
+echo "=== U-sweep: Joint ==="
+for U in 1 2 3 5 6; do
+    echo ">>> Joint U=$U"
+    $PY Code/drl_mask_policy.py \
+        --mode joint --embed-file $EMB \
+        --users $U --mux-factor 4 \
+        --epochs $EPOCHS --steps-per-epoch $STEPS \
+        --save-dir results_sweeps/joint_U${U} \
+        2>&1 | tail -3
+done
+
+echo "=== Ablation: beta=0 (no orthogonality reward) ==="
+$PY Code/drl_mask_policy.py \
+    --mode drl --embed-file $EMB \
+    --users 4 --mux-factor 4 --beta 0.0 \
+    --epochs $EPOCHS --steps-per-epoch $STEPS \
+    --save-dir results_sweeps/drl_beta0 \
+    2>&1 | tail -3
+
+echo "=== Ablation: r=16 (smaller rank) ==="
+$PY Code/drl_mask_policy.py \
+    --mode drl --embed-file $EMB \
+    --users 4 --mux-factor 4 --rank 16 \
+    --epochs $EPOCHS --steps-per-epoch $STEPS \
+    --save-dir results_sweeps/drl_r16 \
+    2>&1 | tail -3
+
+echo "=== Ablation: r=d_s=3072 (full rank) ==="
+$PY Code/drl_mask_policy.py \
+    --mode drl --embed-file $EMB \
+    --users 4 --mux-factor 4 --rank 3072 \
+    --epochs $EPOCHS --steps-per-epoch $STEPS \
+    --save-dir results_sweeps/drl_rfull \
+    2>&1 | tail -3
+
+echo "=== All sweeps complete ==="
+ls -la results_sweeps/